Ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory

Abstract

The invention discloses an ocean cloud detection probability calculation method based on Gaussian mixture distribution and a Bayesian theory, and belongs to the technical field of remote sensing image processing and meteorological monitoring, and the method comprises the steps: reading and preprocessing multichannel data of a stationary meteorological satellite, carrying out the preprocessing, selecting cloud sensitive features, and carrying out the statistics of a double-peak histogram of each feature; initializing a Gaussian mixture model (GMM) by adopting K-means clustering, estimating a cloud prior probability, and fitting the Gaussian mixture model in combination with an EM algorithm to obtain a cloud / non-cloud probability density; and calculating a cloud posterior probability by using a Bayesian theory, carrying out weighted fusion on a multi-feature probability, then carrying out smoothing processing, and outputting a cloud probability field or a binary mask. The method breaks through the limitation of traditional binary classification, quantifies uncertainty through probabilistic output, reduces the misjudgment rate through multi-feature fusion, improves the detection accuracy and flexibility, adapts to different satellite data and marine environments, and provides reliable cloud information for weather forecast and marine parameter inversion.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image processing and meteorological monitoring, and particularly relates to a marine cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory. BACKGROUND

[0002] In marine observation and meteorological prediction, it is of great significance to accurately obtain the cloud distribution information of the marine area. Cloud coverage not only affects the radiation and heat flux of the marine surface, but also directly relates to the monitoring and prediction of meteorological events such as typhoon and low pressure, and affects the remote sensing observation accuracy of marine water color, surface temperature and wind field.

[0003] In the existing marine cloud detection, the commonly used methods are mostly based on threshold segmentation or simple statistical discrimination, which directly divide the pixels into two categories of "cloud" and "non-cloud". Although simple to implement, the following problems exist: first, the threshold division method is too rigid and difficult to adapt to different imaging conditions (such as day and night, seasons) and diversified marine environments (such as tropical and temperate sea areas); second, in the case of overlapping distribution of cloud and non-cloud features, forced binary classification can easily cause high misjudgment rate (such as solar flare area easily misjudged as cloud); third, the binary classification result lacks uncertainty information, which is difficult to support risk assessment in subsequent meteorological model integration or disaster monitoring, and cannot meet the high-precision application requirements. Therefore, a method capable of outputting marine pixel cloud probability is urgently needed to improve the accuracy and reliability of marine cloud detection. SUMMARY

[0004] The purpose of the present application is to solve the problems of rigid classification, high misjudgment rate and lack of uncertainty information in the existing marine cloud detection method. The present application provides a marine cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory, which realizes the transition from traditional binary classification to probabilistic output by outputting continuous cloud probability values, reduces the misjudgment risk of single feature by fusing multiple cloud-sensitive features to improve the detection accuracy in complex environments, and constructs an extensible method framework to adapt to different stationary meteorological satellite data and marine meteorological conditions, thereby providing more accurate and reliable cloud distribution information for marine observation, meteorological prediction and remote sensing observation of marine water color, surface temperature, etc.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] A marine cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory, comprising the following steps:

[0007] S1, data reading and preprocessing: reading the multi-channel observation data of stationary meteorological satellite, performing radiation calibration, geometric correction and quality control on the data, and removing land pixels through sea-land mask, calculating the flare angle Φ and removing flare pollution pixels with Φ < 10°;

[0008] S1, feature selection and histogram statistics: select features sensitive to clouds from the preprocessed multi-channel data, calculate the histogram bandwidth using the Freedman-Diaconis criterion, and statistically analyze the histogram distribution of each feature;

[0009] S3, K-means clustering: K-means algorithm is used for clustering analysis of each feature, and "cloud" and "non-cloud" categories are divided. The proportion of cloud pixels is calculated as the prior probability of cloud, and the initial value is provided for subsequent histogram fitting, so that the fitting starts from a position closer to the true distribution, converges quickly, and also avoids the situation that the optimal solution cannot be found due to improper initial value setting;

[0010] S4, GMM parameter estimation: taking the K-means clustering result as the initial value, based on Gaussian Mixture Model (GMM), using Expectation Maximization (EM) optimization algorithm to estimate the parameters of the histogram, fitting the feature histogram, and obtaining the probability density distribution of cloud and non-cloud (water body);

[0011] S5, Bayesian posterior probability calculation: combining Bayesian theory, according to the prior probability provided by K-means clustering and the probability density distribution provided by Gaussian Mixture Model, the posterior probability of the pixel belonging to cloud is calculated, reflecting the confidence of the pixel being cloud;

[0012] S6, multi-feature fusion: according to the feature discrimination ability, the weight is allocated, the cloud posterior probability obtained by different features is weighted and fused, the comprehensive probability value of each pixel belonging to cloud category is calculated, the cloud probability field is generated, the complementary advantages of various spectral features are fully utilized, and the misjudgment and omission under single feature are avoided;

[0013] S7, probability field smoothing: filtering processing is performed on the cloud probability field, and spatial smoothing operation significantly reduces isolated misjudgment and discontinuous area, and improves the overall stability of the detection result;

[0014] S8, result output: according to the actual needs, the smoothed cloud probability field is output as the final cloud detection result for uncertainty evaluation; or set a threshold for the cloud probability field, output a binary cloud mask; or directly divide the comprehensive probability value into probability intervals, output the probability interval result, and intuitively show the cloud distribution confidence.

[0015] As a further description of the above technical solution, the calculation formula of the flare angle Φ in step S1 is:

[0016] ;

[0017] Where θ is the solar elevation angle, α is the sensor elevation angle, and φ is the relative azimuth angle between the sun and the sensor.

[0018] As a further description of the above technical solution, the cloud-sensitive features in step S2 include infrared channel brightness temperature, visible light channel reflectivity (such as 0.86 μm channel reflectivity R 0.86 ) or a combination feature (such as the infrared brightness temperature difference ΔT of 3.9 μm channel and 11.2 μm channel).

[0019] As a further description of the above technical solution, the calculation formula of the histogram bandwidth in step S2 is:

[0020] ;

[0021] In the formula, N is the number of statistical samples, and IQR is the quartile difference of the samples.

[0022] As a further description of the above technical solution, the probability density distribution of the cloud in step S4 is:

[0023] ;

[0024] The probability density distribution of the non-cloud (water body) is:

[0025] ;

[0026] Wherein, μ cloud and σ cloud are the mean and standard deviation of the cloud, μ water and σ water are the mean and standard deviation of the non-cloud (water body), and x i is the reflectivity of the pixel.

[0027] As a further description of the above technical solution, the calculation formula of the posterior probability in step S5 is:

[0028] ;

[0029] Wherein, P(cloud) is the proportion of cloud pixels, P(water) is the proportion of non-cloud (water body) pixels, P(x i ) is the probability density of the feature marginal distribution, and:

[0030] .

[0031] As a further description of the above technical solution, the calculation formula of the comprehensive probability value in step S6 is:

[0032] ;

[0033] Wherein, n is the number of features, W iFor the weight of each feature, the observation conditions (such as the solar elevation angle, the regional time period, etc.) can be adaptively adjusted to improve the robustness of the method in complex environments.

[0034] As a further description of the above technical solution, the filtering process in step S7 is median filtering, mean filtering or Gaussian filtering.

[0035] As a further description of the above technical solution, the probability interval division in step S8 is: low-probability cloud area (<0.3), medium-probability cloud area (0.3-0.7) and high-probability cloud area (>0.7).

[0036] Compared with the prior art, the present application has the following beneficial effects:

[0037] (1) More flexible probability output: not only provides binary discrimination results of cloud / non-cloud, but also gives cloud probability values (0-1) of each pixel, making the detection results more flexible, and users can define thresholds according to application scenarios (such as meteorological forecast requiring high-confidence cloud area, ocean inversion requiring low-misjudgment clear sky area) to adapt to different needs.

[0038] (2) More accurate multi-feature fusion: make full use of multi-dimensional features such as infrared brightness temperature, visible light reflectivity, brightness temperature difference and texture features, complement each other's advantages through Bayesian probability theory, significantly reduce the misjudgment rate of solar flare and thin cloud, etc. scene, improve detection accuracy, and reduce the risk of misjudgment of single feature;

[0039] (3) Stronger scalability: the method framework supports adding new features (such as texture features) or replacing probability modeling algorithms (such as Dirichlet mixture model), adapting to different static meteorological satellites (such as FY-4A, GOES) data and complex ocean environments (nearshore, polar sea area);

[0040] (4) Higher efficiency of initial value optimization: provide GMM fitting initial value through K-means clustering, avoid loss of optimal solution caused by improper initial value, and improve the convergence speed of EM algorithm by more than 30%. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flowchart of the ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory of the present application.

[0042] Figure 2 is the visible light (left) and infrared cloud image (right) of the research area.

[0043] Figure 3 is a histogram fitting diagram based on a Gaussian mixture distribution model (GMM).

[0044] Figure 4 is a cloud probability diagram based on the Bayesian cloud probability calculation result of a single feature.

[0045] Figure 5 is the cloud probability map after multi-feature weighted fusion and probability field smoothing.

[0046] Figure 6 is the cloud probability map after cloud probability field binarization segmentation. DETAILED DESCRIPTION

[0047] The technical solutions of the present application will be further described in detail below in combination with the drawings and specific embodiments, but do not constitute any limitation on the present application, and any limited number of modifications made by anyone within the scope of protection of the claims of the present application is still within the scope of protection of the claims of the present application.

[0048] A marine cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory, the flowchart is as shown in Figure 1 , comprising the following steps:

[0049] S1, data reading and preprocessing:

[0050] Multi-channel remote sensing data is read from a meteorological satellite (such as Himawari-8), and format analysis, radiation calibration, geometric correction and quality control operations are performed on the data to obtain standardized data that can be used for subsequent analysis;

[0051] And use the sea-land mask file to set the land pixels as invalid values, reduce the subsequent calculation amount;

[0052] Considering that the water body contaminated by the sun glint during the day is very close to the spectral characteristics of clouds and fog in visible, near-infrared, and short-infrared or even mid-infrared bands, which is easy to cause misjudgment, the glint angle Φ (glint angle) is calculated according to the sun elevation angle θ, the sensor elevation angle α and the relative azimuth angle φ (the data is derived from the satellite observation header file), and the sun glint contaminated pixels with Φ < 10° are removed, and the glint angle calculation formula is:

[0053] .

[0054] S2, feature selection and histogram statistics:

[0055] Select features sensitive to clouds from the preprocessed multi-channel data, including infrared channel brightness temperature, visible channel reflectivity or their combination, etc. During the day, visible near-infrared channel reflectivity (such as 0.86 μm channel), mid-infrared and long-infrared channel brightness temperature difference (such as 3.9 μm and 11.2 μm brightness temperature difference) and other features can be selected; at night, infrared window region brightness temperature (such as 11.2 μm channel) can be selected. The Freedman-Diaconis criterion is used to calculate the optimal bandwidth of the histogram, and the histogram distribution of each feature is calculated, and the bandwidth calculation formula is:

[0056] ;

[0057] In the formula, N is the number of statistical samples, and IQR is the interquartile range of the samples.

[0058] S3, K-means clustering:

[0059] The K-means algorithm is used for clustering analysis of each feature (K=2, corresponding to "cloud" and "water"). The K-means algorithm takes the squared error as the standard, and finds the cluster with the highest similarity of intra-cluster samples through iteration for a given sample set. According to the bimodal shape of the feature statistical distribution histogram, the sample set composed of all effective pixels in the study area is clustered into two clusters (cluster 1 and cluster 2) according to the value size, and the mean and standard deviation of the two clusters are calculated. The actual physical properties of clouds and non-clouds (water) are combined to determine the category of each cluster ("cloud" or "non-cloud"). After obtaining the clustering results, the proportion of cloud pixels P(cloud) is calculated as the prior probability estimate of the cloud, and the proportion of water pixels P(water)=1-P(cloud). The mean and standard deviation are used as the initial values for subsequent Gaussian fitting.

[0060] S4, Gaussian Mixture Model (GMM) parameter estimation:

[0061] The sample set composed of all pixels in the study area is considered to be composed of two parts: cloud (cloud) and non-cloud (water). The probability density distribution function of cloud pixels is P(x i │cloud), and the probability density function of non-cloud pixels is P(x i │water). According to the distribution characteristics of the feature histogram fitted by GMM, it is assumed that the probability density of cloud pixels P(x i │cloud) and the probability density of non-cloud pixels P(x i │water) both follow Gaussian distribution, then:

[0062]

[0063] ;

[0064] where μ cloud and σ cloud are the mean and standard deviation of cloud pixels, and μ water and σ water are the mean and standard deviation of non-cloud (water) pixels.

[0065] Through iterative optimization by the Expectation Maximization (EM) algorithm, the parameterized estimates of the mean, variance, and mixing weight of each Gaussian component are obtained. During iteration initialization, the mean and standard deviation of cloud pixels and non-cloud pixels obtained in the K-means clustering step are used as the initial values.

[0066] S5, Bayesian posterior probability calculation:

[0067] Cloud probability estimation relies on the assumption of statistical distribution of pixel features and Bayesian inference, combining prior probability and GMM probability density, to calculate the posterior probability P(cloud | x i ) of each pixel belonging to cloud through Bayesian formula:

[0068] For a certain pixel, we want to determine its category according to its features, i.e. to calculate the posterior probability P(cloud | x i ) of each pixel belonging to cloud, which reflects the confidence of the pixel being cloud, and the formula is:

[0069] ;

[0070] Where P(x i ) is the probability density function of the feature marginal distribution, and:

[0071] .

[0072] S6, multi-feature fusion:

[0073] Assign weights W i to different features, and the weights are determined according to the discriminant ability of the features in distinguishing cloud and non-cloud (e.g. 0.86 μm reflectivity has strong discriminant ability during the day, and brightness temperature difference is more reliable at night); weight the cloud probability results of different features to calculate the comprehensive probability value of each pixel belonging to cloud category, generate the cloud probability field (cloud probability distribution map), and play the complementary advantages of features, and the calculation formula of the comprehensive probability value is:

[0074] ;

[0075] Where n is the number of features, and W i is the weight of each feature, which can be adjusted adaptively according to the observation conditions (such as solar elevation angle, regional time period, etc.) to improve the robustness of the method in complex environments.

[0076] S7, probability field smoothing:

[0077] Perform filtering on the cloud probability field, including median filtering, mean filtering or Gaussian filtering, etc., to reduce isolated misjudgment points and discontinuous regions, and improve the spatial continuity and stability of the probability field.

[0078] S8, result output:

[0079] The smoothed cloud probability field can be directly output as a final cloud detection result, which is used for uncertainty assessment (such as focusing on a high-probability cloud area); in a scene requiring rapid application, a threshold can be set for the cloud probability field to obtain a "cloud / non-cloud" binary classification mask; or the probability interval result can be output by interval division (such as a low-probability cloud area, a medium-probability cloud area, and a high-probability cloud area), so as to intuitively express the uncertainty in the cloud detection result.

[0080] Application examples:

[0081] Taking the AHI sensor data of the Himawari-8 satellite (covering the coastal areas of East China and the surrounding sea areas: longitude 116°E-128°E, latitude 22°N-42°N, spatial resolution 2 km, channels containing 0.86 μm, 3.9 μm and 11.2 μm) as an example, combined with Figures 2-6 The implementation process of the present application is described in detail as follows:

[0082] S1, data reading and preprocessing:

[0083] From the Himawari-8 Level 1B data file, the 0.86 μm (visible light), 3.9 μm (medium-wave infrared), and 11.2 μm (long-wave infrared) channel data of March 28, 2021, 01:00 UTC were read;

[0084] Radiometric calibration: using the scaling coefficients provided by the satellite, the original digital values (Digital Number) are converted into reflectivity (0.86 μm channel) and brightness temperature (3.9 μm, 11.2 μm channels);

[0085] Geometric correction: based on the WGS84 coordinate system, through satellite orbit parameters and ground control points, the latitude and longitude deviations of the data are corrected to ensure the accuracy of the pixel position;

[0086] Sea-land mask: global 1 km resolution sea-land mask data is called to set the land pixels within the latitude range of 22°N-42°N and the longitude range of 116°E-128°E to invalid values;

[0087] Flare removal: the solar elevation angle θ=35°, the sensor height angle α=60°, and the relative azimuth angle φ=45° are calculated, and cosΦ≈0.9848, Φ≈10° is obtained by substituting the flare angle formula, and the pixels with Φ<10° (mainly distributed in the equatorial sea area) are removed.

[0088] The cloud image after preprocessing is as follows: Figure 2As shown, it is divided into two channels, showing the characteristics of the cloud from different dimensions: the left figure is the true color channel, where the white area is the cloud cluster, and the dark color is the ocean or land, which can intuitively identify the macroscopic morphology and coverage of the cloud; the right figure is the infrared 11.2 μm channel brightness temperature map, which reflects the physical characteristics of the cloud through the brightness temperature (radiation temperature of the cloud layer or sea surface), the lower the brightness temperature, the higher and thicker the cloud layer; the higher the brightness temperature, the thinner the cloud layer or the sea surface.

[0089] S2, feature selection and histogram statistics:

[0090] Feature selection: select the reflectivity of the visible light 0.86 μm channel and the brightness temperature difference between the infrared 3.9 μm channel and the infrared 11.2 μm channel as two features;

[0091] Histogram statistics: statistics of the reflectivity characteristics and brightness temperature difference characteristics of the effective pixels (N=50000), the four quarter differences of the reflectivity IQR=0.3 and the four quarter differences of the brightness temperature difference IQR=5 K are calculated by the statistics software / library; the bandwidth of the reflectivity is calculated by the bandwidth formula =2x0.3 / 50000 1 / 3 ≈0.02, and the bandwidth of the brightness temperature difference is =2x5 / 50000 1 / 3 ≈0.33, draw a histogram as Figure 3 shown (gray area), showing a bimodal distribution, corresponding to cloud and sea water respectively.

[0092] S3, K-means clustering:

[0093] R 0.86 characteristics perform K-means clustering (K=2): with the minimum square error minimization as the goal, iteratively divided into two categories, calculate the mean and standard deviation of each category: cluster 1 mean=0.6, standard deviation=0.1 (high reflectivity, determined as cloud), cluster 2 mean=0.1, standard deviation=0.05 (low reflectivity, determined as sea water); the proportion of cloud pixels is 30%, i.e. the prior probability of cloud P(cloud)=0.30, then the non-cloud prior probability P(water)=0.70.

[0094] S4, GMM parameter estimation (EM algorithm):

[0095] Initialization: the cloud / non-cloud mean and standard deviation obtained by K-means clustering are used as the initial parameters (μ cloud =0.6, σ cloud =0.1; μ water =0.1, σ water =0.05) of the Gaussian mixture model (GMM), take a pixel x i =0.5 (0.86 μm visible light channel reflectivity) as an example, calculate the probability density of the cloud and the probability density of the water body:

[0096] Probability density of cloud: ;

[0097] Probability density of water: ;

[0098] E-step: Calculate the posterior probability that the pixel belongs to cloud ;

[0099] M-step: Update the GMM parameters according to the formula , and similarly for μ water , σ water and the mixing weights;

[0100] Iteration: Repeat the E-step and M-step until the parameter change is <10 -4 , (iteration number ~ 15 times), and finally we get μ cloud = 0.62, σ cloud = 0.09; μ water = 0.09, σ water = 0.04.

[0101] S5, Bayesian probability calculation:

[0102] Calculate the probability density of cloud: ;

[0103] Calculate the probability density of water: ;

[0104] Fit the histogram, the results are shown in Figure 3 , the PDF curve is highly consistent with the histogram.

[0105] Combine the prior probability and the probability density to calculate the posterior probability that each pixel belongs to cloud:

[0106] (high probability is cloud)

[0107] The results are shown in Figure 4The left image is a cloud detection VIS image based on the reflectance of the 0.86 μm channel. The color bar ranges from 0 (blue, low cloud probability) to 1 (red, high cloud probability). The red area represents a very high probability of clouds, and the blue area represents almost no clouds. As can be seen from the image, there is a large area of high probability of clouds (red) in the study area, which is highly consistent with the visual form of the actual cloud system (gray and white cloud clusters in the image), indicating that the visible light channel can effectively identify the reflection characteristics of clouds. The right image is a cloud detection IR image based on the brightness temperature difference between the 3.9 μm and 11.2 μm channels. Comparing the left image, the high probability cloud area of the infrared channel is basically consistent with that of the visible light channel, but there are differences in details (such as the identification of thin clouds and low clouds) - this is because the visible light channel is more sensitive to the "reflection strength" of clouds, and the infrared channel is more sensitive to the "temperature level" of clouds (high clouds and thick clouds have low temperatures, and the probability of identification is more significant). The effectiveness of single-feature Bayesian probability calculation is verified, laying a foundation for subsequent multi-feature fusion.

[0108] S6, multi-feature weighted fusion:

[0109] Weight distribution: based on the feature discrimination ability, the discrimination ability of the reflectance of the 0.86 μm channel is strong during the day, and the brightness temperature difference of the infrared 3.9 μm channel and the infrared 11.2 μm channel is used for auxiliary discrimination, and the weights W1=0.6 (reflectance) and W2=0.4 (brightness temperature difference) are set.

[0110] The reflectance of the 0.86 μm channel of a certain pixel corresponds to P1=1.0, and the brightness temperature difference of the infrared 3.9 μm channel and the infrared 11.2 μm channel corresponds to P2=0.95, so the comprehensive cloud probability is:

[0111] .

[0112] S7, probability field smoothing:

[0113] A 3x3 window median filter is used to process the comprehensive probability field: taking a certain pixel as the center, the probability values of the 9 pixels in the 3x3 window are sorted, and the middle value is taken as the smoothed probability of the pixel. Eliminate isolated high probability points (such as flare residual misjudgment), and improve spatial continuity.

[0114] As Figure 5As shown, large areas of the study area are red, which highly matches the visual morphology of actual cloud systems (grayish-white cloud clusters in the image), indicating that these areas have a high probability of being identified as clouds. The blue areas are mainly distributed in open sea areas or at the edges of cloud systems, and are more likely to be seawater or thin / fragmented clouds. Compared with single-channel detection, the weighted fusion of multiple features (such as visible light and infrared channels) provides better spatial continuity and accuracy. It not only quantifies the probability of cloud presence but also intuitively displays the confidence level of clouds in different areas through color gradients, providing accurate cloud information with uncertainties for scenarios such as weather forecasting (e.g., typhoon monitoring) and marine observation.

[0115] S8. Output of Results:

[0116] Direct output Figure 5 The cloud probability field shown is saved in GeoTIFF format, containing latitude and longitude information and probability values ​​(0-1), for use in meteorological model uncertainty assessment.

[0117] Generate a binary mask: Set a threshold of 0.5. Pixels with a probability > 0.5 are labeled "cloud" (value 1), otherwise "non-cloud" (value 0), generating a "cloud / non-cloud" binary mask, such as... Figure 6 As shown in the figure, the red area represents the high-probability cloud area, and the blue area represents the low-probability cloud area. By quantifying the probability of cloud existence, it provides accurate cloud information with uncertainty for weather forecasting (such as typhoon monitoring) and marine observation. The red area has high cloud confidence and is the key area of ​​focus for meteorological analysis.

[0118] The probability intervals are divided (<0.3: low probability cloud area, 0.3~0.7: medium probability cloud area, >0.7: high probability cloud area) to distinguish between sea fog (medium probability area) and thick clouds (high probability area).

[0119] The specific embodiments described herein are merely illustrative of the principles and technical means of the invention. The examples are intended to aid in understanding the method and core ideas of the invention. Furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. A method for calculating the probability of ocean cloud detection based on Gaussian mixture distribution and Bayesian theory, characterized in that, Includes the following steps: S1. Data Reading and Preprocessing: Read multi-channel data from geostationary meteorological satellites, perform radiometric calibration, geometric correction, quality control and land-sea masking operations, calculate the flare angle Φ and remove pixels with Φ < 10°; S2. Feature selection and histogram statistics: Select cloud-sensitive features from the preprocessed multi-channel data, calculate the histogram bandwidth using the Freedman-Diaconis criterion, and statistically analyze the histogram distribution of each feature. S3, K-means clustering: The K-means algorithm is used to perform cluster analysis on each feature, divide the category into "cloud" and "non-cloud", and count the proportion of cloud pixels as the prior probability of cloud, and provide initial values ​​for subsequent fitting; S4. GMM parameter estimation: Using the K-means clustering results as initial values, the feature histogram is fitted based on the Gaussian mixture model and the expectation-maximization optimization algorithm to obtain the probability density distribution of clouds and the probability density distribution of non-clouds. S5. Bayesian probability calculation: Calculate the posterior probability that a pixel belongs to the cloud based on the prior probability provided by K-means clustering and the probability density distribution provided by Gaussian mixture model. S6. Multi-feature weighted fusion: Weights are assigned according to the feature discrimination ability, and the cloud posterior probabilities of different features are weighted and fused to calculate the comprehensive probability value of each pixel belonging to the cloud category and generate a cloud probability field. S7. Probability Field Smoothing: Performs filtering on the cloud probability field to reduce isolated misjudgments and discontinuous areas; S8. Result Output: Directly output the cloud probability field for uncertainty assessment; or generate a binary cloud mask by setting a threshold according to user needs; or intuitively display the cloud distribution confidence by dividing the probability interval.

2. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that, The formula for calculating the flare angle Φ in step S1 is: ； Where θ is the solar altitude angle, α is the sensor altitude angle, and φ is the relative azimuth angle between the sun and the sensor.

3. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that: The cloud-sensitive features mentioned in step S2 include infrared channel brightness temperature, visible light channel reflectance, or a combination of these features, wherein the combination of features includes brightness temperature difference.

4. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1 or 3, characterized in that, The formula for calculating the histogram bandwidth in step S2 is as follows: ； In the formula, N is the number of statistical samples, and IQR is the interquartile range of the sample.

5. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that, The probability density distribution of the cloud in step S4 is as follows: ； The probability density distribution of the non-cloudy component is as follows: ； Where, μ cloud and σ cloud Let μ be the mean and standard deviation of the cloud. water and σ water Here, x represents the mean and standard deviation of the non-cloudy data, respectively. i denoted as the reflectance of a pixel.

6. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that, The formula for calculating the posterior probability mentioned in step S5 is as follows: ； Where P(cloud) is the proportion of cloud pixels, P(water) is the proportion of non-cloud pixels, and P(x) is the proportion of non-cloud pixels. i Let be the probability density of the characteristic marginal distribution, and: 。 7. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that, The formula for calculating the comprehensive probability value mentioned in step S6 is as follows: ； Where n is the number of features, W i The weights of each feature.

8. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that: The filtering process described in step S7 is median filtering, mean filtering, or Gaussian filtering.

9. The ocean cloud detection probability calculation method based on Gaussian mixture distribution and Bayesian theory according to claim 1, characterized in that: The probability intervals mentioned in step S8 are divided as follows: a probability < 0.3 is a low probability cloud area, a probability between 0.3 and 0.7 is a medium probability cloud area, and a probability > 0.7 is a high probability cloud area.

Images